基于小样本及贝叶斯推理的避雷器缺陷分类技术

Classification of arrester defects based on small sample data and Bayesian inference

不同设备和缺陷发生概率的差异性导致避雷器缺陷样本具有小样本特点,难以训练准确的预测模型。仅依靠避雷器监测指标及阈值判断缺陷,很难识别出避雷器监测数据未超限值平稳增长时的异常状态。因此提出一种基于小样本数据及贝叶斯推理的避雷器缺陷分类方法,扩展数据维度,建立多阶段信息的缺陷诊断模型,利用贝叶斯推理算法,计算不同特征量下缺陷原因类别的概率。根据新的检测证据更新不同特征量下缺陷原因类别的概率,从而识别缺陷原因。算法基于新的检测数据和标注结论,自动调整模型中关于设备缺陷和原因的先验概率指标,保证缺陷原因分类准确率。利用某电力公司避雷器运维数据和在线监测系统,分析和验证了本文所提方法的有效性和正确性,为避雷器运维提供有效支撑。

Due to the difference of occurrence probability of different devices and defects,the samples of arrester defects have the characteristics of small samples,so it is difficult to train an accurate prediction model.It is difficult to identify the abnormal state when the monitoring data of arrester does not exceed the limit value and grow steadily only depending on the monitoring index and threshold value of arrester.Therefore,a method of arrester defect classification based on small sample data and Bayesian reasoning is proposed.On the basis of on-line monitoring,the data dimension is expanded,and a defect diagnosis model with multi-stage information is established.Bayesian reasoning algorithm is used to calculate the probability of defect cause categories under different feature quantities.According to the new detection evidence,the probability of defect cause categories under different feature quantities is updated to identify the defect causes.Based on the new detection data and labeling conclusion,the algorithm automatically adjusts the prior probability index of equipment defects and causes in the model to ensure the accuracy of defect cause classification.The effectiveness and correctness of the proposed method are analyzed and verified by using the operation and maintenance data and on-line monitoring system of a power company,which provides effective support for the operation and maintenance of arresters.